Reading the paper and looking at you're implementation, this is certainly in 
the ball park I am striving for. The way I think of it is each ''spreadsheet 
cell'' should look after itself, it's called data flow architectures in some of 
the older literature. 

My current implementation uses Iterators and a my data is split over several 
column qualifiers, which I know will be processed in order. Actions on current 
column are depend on the state of previous columns. What I'm trying to avoid 
are disk seeks - if I can fold updates in during compaction I can reduce wasted 
operations. 

I've effectively got context for the Observer. 

Tnx for the deadlock scenarios - I'm pretty certain it is Situation 1.

Peter


________________________________
 From: Keith Turner <[email protected]>
To: [email protected]; Peter Tillotson <[email protected]> 
Sent: Monday, 15 July 2013, 13:49
Subject: Re: Iterators - updating other rows
 







On Mon, Jul 15, 2013 at 6:38 AM, Peter Tillotson <[email protected]> wrote:

I've got two tables of dependent data, which I was hoping to update efficiently 
during compaction. This leads to the following requirements:
>  - Changes to other rows
>  - Changes in other tables
>
>
>I've fought with iterators and embedding writers, but have had to fall back to 
>map reduce jobs to complete the update. 
>
>
>Is there a recommended approach to this?

Writing to Accumulo from an iterator can lead to deadlock.  I can think of at 
least the following two situations, but there are probably more.

Situation 1 

 1. Memory is full on tablet server 1 and writes are held
 2. Tablet X is on Tserver 1 and is scheduled for compaction to free memory
 3. Tablet X tries to write to Tablet server 1, but the writes block because 
memory is full (deadlock)
 4. No other tablet on Tserver 1 can be written to because memory is full and 
can not be flushed, 
     so the problem snowballs


Situation 2

 1. Tserver 2 is hosting Tablet Y & Z
 2. Tablet Y & Z have data in memory
 3. Tserver 2 dies
 4. Tserver 3 loads Tablet Y, recovers its data, and tries to compact
 5 Tablet Y tries to write to Tablet Z during compaction 
 6. Tserver 4 loads Tablet Z, recovers its data, and tries to compact
 7 Tablet Z tries to write to Tablet Y during compaction 
 8. Tablets Y & Z are not loaded yet, but trying to write each other (deadlock)
 9. Tablet servers 2 and 3 can not load any more tablets, because their load 
threads are both stuck.
     so the problem snowballs

I am currently working on an implementation of Percolator[1].  Not something 
you can use now, but I am curious if you could use Percolator to solve your 
problem?  I am very interested in feedback on this project while its in its 
formative stages.  I hope to have it finished w/ Accumulo 1.6.0.

[1]: https://github.com/keith-turner/Accismus


>
>I bit more detail about the algorithm. 
>
>
>I've two tables with different sort orders, and I use ngram row ids to group 
>element and split over multiple tablets, so:
>
>
>Table1
>nm: key1: 000: newValueId2
>nm: key2: type: valueId1
>nm: key3: type: valueId1
>
>
>Table2
>ab: valueId1: 001: blob
>ab: valueId1:key2: nm
>..
>..
>    
>Multiple keys point to the same value in the other table but both keys and 
>values are liable to changes ... what I was trying to do was use special 
>columns (column Qaulifier 000 above), I call them care-of to do redirects as 
>data changes real-time, with iterators this would becomes eventually 
>consistent and be very efficiently but a MapReduce approach requires multiple 
>table scans of each large table. I like the approach because the ngram splits 
>/ groups data and the two different sorts give me different nice query 
>characteristics.
>
>
>For some reason the embedded writers were blocking - I may retry with a larger 
>cluster. I fought with it for a few days then resorted to MapReduce jobs until 
>I get a chance to look at the Accumulo code more closely. 
>
>
>Would it be easy to add a special iterator that accepts (Text, Mutation) pairs 
>much as the AccumuloOutputFormat does ?  
>
>
>Many thanks in advance
>
>Peter.

Reply via email to